Mira Murati’s startup ships a 975B open model

Thinking Machines Lab, the startup founded in early 2025 by former OpenAI CTO Mira Murati, has unveiled Inkling, a 975 billion-parameter open-weights model aimed at developers who want a frontier-scale alternative to offerings from Chinese labs.

The company says Inkling is the largest American open-weights model to date. At native 16-bit precision, it needs more than two terabytes of GPU memory to run — roughly eight Nvidia B300 accelerators or sixteen H200s. For smaller deployments, Thinking Machines has also released an NVFP4 quantized version that can run on half as many GPUs.

Thinking Machines positions Inkling as a flexible model for building AI apps as well as general-purpose uses such as chatbots. It is being released under an Apache 2.0 license, allowing users to fine-tune it for their own workloads, and the company’s Tinker platform includes tools for customization. The startup also claims the model can write its own fine-tuning scripts to adjust behavior, learn new skills, and evaluate itself.

Inkling includes a 1 million-token context window, which should help with large codebases and search tasks that require finding small details in large amounts of data. Its mixture of experts design was inspired by DeepSeek-V3, according to the company, but Thinking Machines says it trained the model from scratch on Nvidia GB300 NVL72 systems using 45 trillion tokens of text, images, audio, and video.

Recommended reading

Hassabis says STEM makes you 10x better at AI

Model architecture and performance claims

The model has 256 routed experts and two shared experts. Each generated token uses six experts, for about 41 billion active parameters per token. Thinking Machines says that lets Inkling produce tokens at about the same rate as DeepSeek V4 on the same hardware.

Like many current large language models, Inkling is a reasoning model trained with reinforcement learning to use chain-of-thought-style reasoning before answering. Thinking Machines says it tuned the model to use those thinking tokens more efficiently, claiming Inkling matches Nvidia Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly a third the tokens. The Register notes that benchmark claims should be treated cautiously, and the charts shared by Thinking Machines still show Inkling trailing proprietary models from Anthropic and OpenAI.

Availability on Tinker and Hugging Face

Inkling is available now through the Tinker platform. Thinking Machines says it is also working to bring the model to third-party API providers including:

• TogetherAI
• Fireworks
• Modal
• Databricks
• Baseten

For users who want to test it themselves, the model is also available on repositories including Hugging Face. At launch, Thinking Machines says it supports inference engines such as vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp.

The company is also previewing Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters for users who care more about latency than throughput and quality. Thinking Machines said it is still finalizing Inkling and plans to release the model weights once testing is complete.